Scene Text Detection Based on Multi-Headed Self-Attention Using Shifted Windows

Scene text detection has become a popular topic in computer vision research. Most of the current research is based on deep learning, using Convolutional Neural Networks (CNNs) to extract visual features images. However, due limitations convolution kernel size, CNNs can only local images with small perceptual fields, and they cannot obtain more global features. In this paper, improve accuracy scene detection, feature enhancement module added model. This acquires an image by computing multi-headed self-attention map. The improved model extracts CNNs, while extracting through module. extracted both these are then fused ensure that at different levels extracted. A shifted window used calculation self-attention, which reduces computational complexity from second power input width-height product first power. Experiments conducted multi-oriented dataset ICDAR2015 multi-language MSRA-TD500. Compared pre-improvement method DBNet, F1-score improves 0.5% 3.5% MSRA-TD500, respectively, indicating effectiveness improvement.